This paper describes the parallel implementation of the TRANSIMS traffic micro-simulation.

In this paper, we explain a parallel implementation of a traffic micro-simulation model based on the queue model by Gawron. Within this model, links do not only have a flow capacity that limits the number of vehicles that can leave the link, but also have a space constraint which limits the number of vehicles that can be on a link simultaneously. The vehicles in this model follow their precomputed paths as in other route-plan-based simulations such as TRANSIMS. Since the queue model needs less data and computing power, it runs much faster than TRANSIMS for the same data. In the parallel implementation of the queue model we distribute the data onto a number of processors, each of which runs a smaller portion of the simulation. The parallel version gives a maximum speedup of 28 on 64 CPUs using Myrinet communication. For a realistic simulation of all of Switzerland, computing is up to 190 times faster than real time. The maximum number of vehicles simultaneously driving in that simulation is about 160 000.

We use the so-called queue model introducted by Gawron as the base of the traffic dynamics in our micro-simulation. The queue model describes the links with a flow capacity that limits the number of agents that can leave the link and a space constraint which defines the limit of the number of agents that can be on a link at the same time. Free flow speed is the third key component of traffic dynamics in the model. Flow capacity and space constraint together model physical queues, which can spill back beyond the end of the link. A consequence of this is that fairness between the incoming traffic streams becomes an issue, since in a spill-back situation they cannot be served at their full rate. We implement and verify a simple solution to this; the solution is much simpler than the one chosen in many other models. The traffic micro-simulation is “large-scale ” which means the simulation is capable of modeling the behavior of millions of agents simultaneously. We utilize a parallel implementation to speed up the computation. In this implementation, the data is distributed onto a number of computing node, each of which runs a smaller portion of the data. Data distribution and communication

Abstract not found

Abstract not found